Cardiomyocyte nuclear ploidy is an important indicator of heart development, remodeling, and repair. However, measuring ploidy directly in heart tissue is difficult because the myocardium contains many cell types arranged in a complex structure. Existing methods often require cardiomyocyte-specific immunostaining and labor-intensive three-dimensional image analysis, which limits their use in large-scale studies.
Researchers from Fuwai Hospital, Chinese Academy of Medical Sciences, and the Institute of Software, Chinese Academy of Sciences, have developed CardioNVT, an artificial intelligence platform for high-throughput cardiomyocyte ploidy assessment in situ.
CardioNVT first uses a UNet++-based image segmentation model to automatically identify and segment cardiomyocyte nuclei from DAPI-stained images alone. In adult mouse heart tissue, the platform accurately recognized and segmented cardiomyocyte nuclear regions, achieving high intersection over union and AUC. The model also maintained stable performance in transverse aortic constriction-induced cardiac hypertrophy mice, suggesting that CardioNVT can be applied not only to healthy cardiac tissue, but also to pathological cardiac remodeling. The research team further used Grad-CAM++ to explore the key image features that contributed to the model's decisions. The results showed strong activation in cardiomyocyte nuclear regions, suggesting that CardioNVT identifies nuclei by recognizing information related to nuclear staining patterns.
After cardiomyocyte nuclei are identified, CardioNVT uses a cross-plane tracking strategy based on Segment Anything Model 2 to match the same nucleus across different planes in z-stack images. This allows the platform to reconstruct the three-dimensional volume of each cardiomyocyte nucleus. The research team also designed a post-processing workflow to correct possible tracking errors, broken tracks, and noise-derived signals. The tracking and reconstruction results showed that the nuclear volume distribution generated by CardioNVT displayed a bimodal pattern consistent with previous studies. Based on the previously fluorescence in situ hybridization-validated relationship between nuclear volume and ploidy, researchers can infer cardiomyocyte nuclear ploidy from the reconstructed three-dimensional nuclear volumes.
Importantly, by leveraging nuclear-morphology cues, a source of information largely neglected by earlier segmentation approaches, CardioNVT highlights the feasibility of nuclear feature-based cell type classification, and thus offers a new perspective on the functional refinement of future cell segmentation algorithms. The research team noted that CardioNVT has thus far been developed and validated in mouse heart tissue, but its modular architecture provides substantial potential for extension to other sample types. In the future, through transfer learning and targeted fine-tuning, the platform could be adapted for use in human tissues, formalin-fixed paraffin-embedded samples, or other staining systems, and further expanded to applications such as the identification of specific cellular subpopulations.
